A lossy audio compression algorithm relies on quirks of human hearing. Those who designed it used their precise understanding of how human hearing works. How would you go about designing a lossy audio compression algorithm without being allowed to learn anything about human hearing?
I believe it would be very much impossible. Humans have no intuitive understanding of the details of human hearing and there are an huge amount of possible ways of reducing the size of an audio file (like throwing away every second bit). There is no possible way you can A/B test all the possible ways and humans have no intuitive understanding of the details of hearing. You can’t intuitively decrease the search space. Aethstetics don’t help you, an understanding of how humans actually work is necessary.
Not everything is like lossy audio compression but I believe there are parallels. Aethstetics do sometimes help you decrease the search space and A/B tests get you the rest of the way but my suspicion is that a true understanding of how humans actually work would always be preferable.